diffusion

Introduces Diffusion Semi-Relaxed Fused Gromov-Wasserstein (DsrFGW), a novel method that integrates node features and structural connectivity for graph comparison via optimal transport and diffusion processes, demonstrating improved robustness to noise and missing edges on synthetic tasks.

The paper introduces Nemotron-Labs-Diffusion, a tri-mode language model that unifies autoregressive, diffusion, and self-speculation decoding, achieving superior throughput and efficiency compared to existing models.

PixWorld presents a unified pixel-space diffusion approach for 3D scene reconstruction and generation, overcoming limitations of latent-space methods by using direct image-level supervision and geometry-aware feature alignment. The method outperforms prior generation methods and matches state-of-the-art reconstruction methods.

SeFi-Image is a new open-source text-to-image foundation model built on Semantic-First Diffusion, available in 1B, 2B, and 5B parameter sizes, trained on just 125K A800 GPU hours.

Set Diffusion introduces a new class of language models that interpolates between autoregressive and diffusion models by factorizing token generation over flexible-position, flexible-length token sets. This enables faster decoding and flexible token ordering, achieving better speed-quality tradeoffs on reasoning, summarization, and unconditional generation tasks.

NVIDIA proposes a method to convert any LLM into a faster one by splitting it into two copies: one frozen for context, the other trained to generate multiple tokens in parallel, achieving 2.4x speedup with ~99% quality retention using only 8% of training data.

This paper proposes Multi-Block Diffusion Language Models (MBD-LMs), extending single-block diffusion to concurrent multi-block decoding with improved training strategies like Multi-block Teacher Forcing and an optimized Block Buffer decoding algorithm. Experiments show increased tokens per forward pass and improved accuracy on benchmarks.

BlockPilot proposes an instance-adaptive policy that predicts the optimal block size for diffusion-based speculative decoding, achieving significant speedup with minimal overhead.

Introduces Speculative Refinement (SpecRef), a training-free hybrid decoding strategy that warm-starts a masked diffusion language model from an autoregressive draft using entropy-guided selective masking. Evaluated across six benchmarks, it reveals that code benchmarks conflate structural discovery with logical correctness, identifies a refinement tension phenomenon, and shows that evaluation protocols can produce different model rankings.

NVIDIA released Nemotron-TwoTower-30B-A3B-Base-BF16, a diffusion-based language model that uses block-wise autoregressive diffusion to generate text by iterative denoising of token blocks, achieving 2.42× the generation throughput of the autoregressive baseline while retaining 98.7% of benchmark quality.

This paper introduces an atomistic language model that integrates a 3D atom encoder, Qwen LLM, and diffusion crystal generator to natively handle multimodal materials data, achieving state-of-the-art crystal structure prediction and de novo generation.

This paper introduces DigenRL, a disaggregated RL framework for diffusion-based generative LLMs that uses generation-axis pipeline parallelism and trainer-assisted generation to improve throughput by 1.56-2.10x over existing systems.

This paper introduces Approximate Structured Diffusion, a method that combines conditional random fields (CRFs) with discrete diffusion for sequence labelling. It uses a CRF conditioned on noisy label sequences and approximate mean-field inference, achieving a 16.5% error reduction on POS tagging.

JanusMesh is a fast, training-free framework that generates text-driven 3D visual illusions—a single mesh revealing different semantics from different viewing angles—by decoupling generation into cross-space dual-branch denoising and view-conditioned texture synthesis, achieving high realism in just 3-5 minutes.

MiniT2I is a minimalist direct-RGB text-to-image generator using a pixel-space MM-JiT denoiser with flow matching and frozen FLAN-T5-Large text tokens, with open-source JAX/Flax and PyTorch implementations released along with checkpoints.

Moebius is a 0.22B parameter image inpainting framework that rivals 10B-level models like FLUX.1-Fill-Dev, achieving over 15x faster inference through novel local-global interaction blocks and adaptive distillation strategies.

RepFusion introduces a method to use pretrained multimodal LLMs as noisy representation encoders in diffusion transformers for text-to-image generation, outperforming baselines with similar compute.

The author launches a weekly Video Model Journal Club covering video generation, world models, physical reasoning, diffusion, flow matching, etc. The first in-person talk will be by Yilun Du on Embodied Reasoning with World Models.

This paper introduces SP³, a method using Spherical Encoder priors for Plug-and-Play image restoration, achieving perceptual quality comparable to zero-shot diffusion priors while being 3–630× faster across tasks.

MoVerse generates real-time interactive video from single images by creating 360° panoramas and 3D Gaussian scaffolds, enabling efficient rendering through diffusion-based techniques.